Pith. sign in

REVIEW 4 major objections 6 minor 22 references

YOLO-FireAD: Efficient Fire Detection via Attention-Guided Inverted Residual Learning and Dual-Pooling Feature Preservation

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read YOLO-FireAD claims a 1.8-point mAP gain over YOLOv8n with 51.8% fewer parameters.

desk verdict A lighter YOLO with a real efficiency win, but the accuracy gain is fragile and the prose contradicts its own recall table. read the letter →

arxiv 2505.20884 v3 pith:V35Z56FP submitted 2025-05-27 cs.CV

Weichao Pan , Bohan Xu , Xu Wang , Chengze Lv , Shuoyang Wang , Zhenke Duan , Zhen Tian This is my paper
classification cs.CV
keywords FiredetectionYOLOAttentionmechanismInvertedresidualDualpoolingSmallobjectEfficientneuralnetworksReal-time
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

YOLO-FireAD claims that replacing YOLOv8n's residual blocks and pooling downsampling with an attention-guided inverted residual block (AIR) and a dual max-average pooling fusion block (DPDF) raises fire-detection accuracy while cutting parameters roughly in half. On the fire_detection dataset it reports 34.6% mean average precision at IoU thresholds 50-95 (mAP50-95), versus YOLOv8n's 32.8%, and 35.5% mAP75, with 1.45 million parameters (51.8% fewer than YOLOv8n) and 4.6 GFLOPs (43.2% lower). The aim is to reduce false alarms from fire-like interference and missed small flames without sacrificing real-time efficiency. A sympathetic reader would care because the result suggests fire early-warning systems could run on lighter hardware with no accuracy sacrifice.

What carries the argument

The two load-bearing modules are AIR and DPDF. AIR (Attention-guided Inverted Residual Block) is an inverted residual block whose expansion is a depthwise separable convolution followed by convolutional additive self-attention: features split into Q, K, V, each refined by a gated spatial branch and a channel-reweighting branch, then merged as $\text{DepthwiseConv}_{3\times3}((\hat{Q} + \hat{K}) \odot V)$, with a 0.25 channel-reduction rate. DPDF (Dual Pool Downscale Fusion Block) is a downsampling block that runs 2x2 max pooling and 2x2 average pooling in parallel, passes each through partial convolution (PConv) plus spatial and channel attention, and fuses them with one learnable coefficient: $X_{\text{fusion}} = \alpha X''_{\max} + (1-\alpha) X''_{\text{avg}}$. AIR does the heavy lifting on parameter reduction (about 39% vs the baseline), while DPDF preserves flame-edge and smoke-diffusion detail during resolution reduction, and their interaction is what the paper claims produces 34.6% mAP50-95 at 4.6 GFLOPs.

What would settle it

Retrain YOLOv8n and YOLO-FireAD on the same data split with identical hyperparameters (e.g., AdamW, learning rate 0.002, batch size 128, 200 epochs) across several seeds; if YOLOv8n reaches or exceeds 34.6% mAP50-95, the reported gain is a training-recipe artifact rather than an architectural benefit.

Watch

Extended reading notes

Core claim

The paper's central claim is that two architectural changes to the YOLOv8n backbone and neck deliver both higher accuracy and lower cost on fire detection. AIR embeds a lightweight channel-spatial attention mechanism inside inverted residual blocks, so the network can amplify flame-core responses and suppress reflections and glare; it accounts for most of the parameter reduction. DPDF replaces plain downsampling with a learnable fusion of 2x2 max and average pooling, preserving flame-edge highlights and smoke continuity so small fires survive resolution reduction. Together they reach 34.6% mAP50-95 (1.8 points over YOLOv8n), the best mAP75 among twelve compared baselines, and the smallest model at 1.45M parameters and 4.6 GFLOPs. The paper also reports the best mAP50-95, 8.12%, when all models are tested on a second fire dataset.

Load-bearing premise

The claimed accuracy edge assumes every compared model was trained under the same optimally tuned protocol, but the paper reports only that all algorithms ran in the same computational setup, without per-model hyperparameters, seeds, or test splits.

Editorial extensions

If this is right

  • If the reported numbers hold, YOLO-FireAD gives a roughly 52% parameter reduction and 43% FLOPs reduction over YOLOv8n while improving mAP50-95 by 1.8 points, easing deployment on edge devices.
  • The DPDF dual-pooling fusion should specifically help small and occluded fire instances; the paper attributes its mAP50 improvement of about 1.7% to this mechanism.
  • mAP75 of 35.5% is 1.3 to 5.5 points higher than the compared mainstream YOLO variants, indicating the gain persists at stricter IoU thresholds used to assess localization quality.
  • On a second dataset, YOLO-FireAD again reports the highest mAP50-95 at 8.12%, suggesting the modules generalize rather than overfit a single split.
  • The ablation shows both modules are needed: AIR alone cuts parameters but slightly lowers mAP50-95 to 32.6, while DPDF alone raises it to 34.5; the combination reaches 34.6 with the fewest parameters.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Inference: because DPDF's max-average fusion is not fire-specific, the architecture could transfer to other small-object detection tasks such as smoke, drones, or potholes, but the paper only tests fire datasets.
  • Inference: the ablation suggests DPDF is the accuracy driver while AIR is the efficiency driver, so a testable extension is making the DPDF fusion coefficient scale-dependent or learned per channel rather than a single global weight.
  • Inference: since only one training protocol is reported, an equally tuned YOLOv8n baseline could close part of the gap; the paper's own ablation shows AIR alone slightly hurts mAP50-95, so the headline gain depends on the pairing of both modules.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. This manuscript proposes YOLO-FireAD, a YOLOv8-based fire detection model whose two main contributions are the Attention-guided Inverted Residual Block (AIR) and the Dual Pool Downscale Fusion Block (DPDF). The paper reports experiments on two Paddle-derived datasets (fire_detection and fire) and claims that YOLO-FireAD reaches 34.6 mAP50-95 on fire_detection, which is 1.8 points above YOLOv8n, while reducing parameters from 3.01M to 1.45M. A cross-dataset test and an ablation study are also reported. The central efficiency claim is plausible, but the accuracy claim is complicated by internal contradictions in the reported tables.

Significance. The efficiency numbers are the strongest part of the paper: a 51.8% parameter reduction and 43.2% FLOPs reduction relative to YOLOv8n with a small mAP75 gain would be practically useful for edge fire detection. The authors make a good-faith effort at cross-dataset validation, which is more than many detection papers do. However, the headline 'outperforms YOLOv8' is not yet supported by the evidence as presented: the model's recall is the lowest in Table 1, its mAP50-95 is below YOLOv12n, and the ablation table does not show that both modules contribute positively to accuracy. These issues are fixable with additional experiments and more careful claims, so the work is of moderate significance at this stage.

major comments (4)
  1. [Section 4.1, Table 1] The prose in Section 4.1 states that YOLO-FireAD's P, R, F1, and mAP are 'significantly better than the other models,' but Table 1 shows the opposite for several headline metrics: recall is 53.3, the lowest of all 13 rows; mAP50 is 64.9, below YOLOv12n's 66.1; and mAP50-95 is 34.6, below YOLOv12n's 35.2. This contradicts the abstract and Section 5 conclusion that the model 'outperforms YOLOv8 by 1.8% mAP' and the 'good precision-efficiency balance' framing. Please restrict the outperformance claim to the metrics that actually support it (mAP75, precision, GFLOPs, parameters) and explicitly discuss why the precision gain comes with the lowest recall.
  2. [Section 3.2, Table 1] The comparison uses one fixed training recipe for all models (batch 128, AdamW lr 0.002, momentum 0.9, 200 epochs) with no per-baseline hyperparameter adjustment, no number of seeds, and no train/validation split definition. Because the claimed mAP50-95 advantage over YOLOv8n is only 1.8 points, which is the same order as seed-to-seed variation on small datasets, the current Table 1 cannot distinguish an architectural benefit from training-recipe or initialization luck. Please report mean and standard deviation over at least three seeds, specify the split, and justify the single-recipe protocol for baselines whose standard training recipes differ (e.g., SGD for YOLOv8n).
  3. [Table 3] The ablation data do not support the statement that 'both AIR and DPDF play an indispensable role in YOLO-FireAD.' Relative to the YOLOv8n baseline, AIR alone decreases mAP50-95 from 32.8 to 32.6 and mAP50 from 64.2 to 63.0; the full model has mAP50 64.9, lower than DPDF-only 65.9, and its recall collapses from 63.8 to 53.3. The paper should either reinterpret AIR as primarily a parameter-reduction mechanism (which Table 3 does show) or provide evidence that the accuracy loss is compensated elsewhere; as written, the module-level contribution claims in Section 2 and Contribution 2 are not supported by the ablation.
  4. [Section 4.1, Table 2] The generalization claim based on the fire dataset is also only partially supported. YOLO-FireAD reaches the best mAP50-95 (8.12), but its mAP50 (25.0) is below YOLOv8n (25.6) and YOLO11n (27.2), and its recall (25.7) is among the lowest in the table. The near-tie with YOLOv12n (8.00) should be discussed, as should the consistency of the single-run protocol applied to this second dataset.
minor comments (6)
  1. [Abstract and Section 4.1] The abstract contains the typo 'YOL-Ov9t' for 'YOLOv9t', and Section 4.1 refers to 'mAP50-90' where 'mAP50-95' is meant.
  2. [Section 4.2, Table 3] The text says the combined model's Model Size is 6.1MB, but Table 3 reports 3.3 for the full model; the numbers should be reconciled.
  3. [Section 2.2, Eq. (3)] Equation (3) uses X after Eq. (2) introduced Xdw; state explicitly that X in Eq. (3) is Xdw, or rename the variable.
  4. [Section 2.3, Eq. (11)] The slicing notation X[r::] in Eq. (11) is ambiguous without a channel-dimension specification; please define the split explicitly.
  5. [Section 2.2 and Section 2.3] The hyperparameters r=0.25 in AIR and r=4 in DPDF are described as 'experimentally optimised'; clarify whether these were selected on the same test set, since that would inflate the reported numbers.
  6. [Section 6] The acknowledgement states that this is a corrected version of an ICIC 2025 paper; the relationship to that prior publication should be stated in the introduction, not only in the acknowledgement.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: reported gains are direct benchmark measurements, not derived from assumed equations.

full rationale

The paper's central claims are empirical measurements summarized in Tables 1-3; the headline comparisons (34.6 vs 32.8 mAP50-95 and 1.45M vs 3.01M parameters) are arithmetic relations on those reported numbers, not quantities obtained by deriving a result from an equation that already contains it. The AIR and DPDF modules are constructive architectural proposals with stated hyperparameters; the text explicitly says r=0.25 was 'experimentally optimised' (Sec. 2.2), which is hyperparameter selection rather than a fitted parameter being renamed as a prediction. The self-citations (refs. [3] and [7]) appear only in background discussion of existing YOLO-based solutions and attention/multi-scale attempts; they are not used as a load-bearing uniqueness theorem or as the sole justification for the proposed architecture. The cross-dataset test in Table 2 provides an external check. Under the specified circularity rubric, there is no step where a prediction reduces by construction to its own input, so the appropriate finding is no significant circularity.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The claims rest on the two datasets, the fairness of the baseline comparison, and the correctness of the cited attention and partial convolution primitives. These are domain assumptions rather than derived statements. The model also relies on two hand set reduction rates and a learned fusion coefficient.

free parameters (3)
  • AIR channel reduction rate r = 0.25
    Chosen by hand and described as experimentally optimised in Section 2.2 (Eq. 1); it controls the information bottleneck of the AIR module and affects the accuracy and efficiency tradeoff.
  • DPDF channel reduction rate r = 4
    Set in Section 2.3 (Eq. 11) to reduce FLOPs; the choice is justified by a claimed 68% FLOPs cut and 90% feature accuracy, but it is still a hand set constant.
  • DPDF fusion weight alpha = not reported
    Learnable coefficient in Eq. 14 fusing max and average pooling paths. It is fit to each dataset during training and its final values are not reported, so the reported accuracies depend on this fitted scalar.
assumptions (3)
  • domain assumption The fire_detection and fire datasets, described only as from the Paddle community, have fixed train and test splits and reliable fire annotations.
    Section 3.1 gives no dataset statistics, split information, or labeling protocol, yet every reported mAP depends on these splits.
  • domain assumption All baseline models were trained under identical, comparably tuned settings.
    Section 3.2 says all algorithms were operationally consistent and ran in the same computational setup, but no per model training details, epochs, or tuning budgets are given, so the comparison assumes fairness.
  • standard math Convolutional additive self attention (CAS-ViT) and partial convolution behave as described in refs 11 and 12 and can be reimplemented as claimed.
    The AIR and DPDF blocks directly embed these prior methods (Sections 2.2 and 2.3); if the cited implementations differ, the reported numbers would change.

how reviews work

0 comments
Cite this review

Pith. "Pith review of YOLO-FireAD: Efficient Fire Detection via Attention-Guided Inverted Residual Learning and Dual-Pooling Feature Preservation." pith.science (2026). https://pith.science/paper/V35Z56FP

@misc{pith2026250520884,
  author       = {Pith},
  title        = {Pith review of: YOLO-FireAD: Efficient Fire Detection via Attention-Guided Inverted Residual Learning and Dual-Pooling Feature Preservation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/V35Z56FP}},
  note         = {Machine review of arXiv:2505.20884}
}
read the original abstract

Fire detection in dynamic environments faces continuous challenges, including the interference of illumination changes, many false detections or missed detections, and it is difficult to achieve both efficiency and accuracy. To address the problem of feature extraction limitation and information loss in the existing YOLO-based models, this study propose You Only Look Once for Fire Detection with Attention-guided Inverted Residual and Dual-pooling Downscale Fusion (YOLO-FireAD) with two core innovations: (1) Attention-guided Inverted Residual Block (AIR) integrates hybrid channel-spatial attention with inverted residuals to adaptively enhance fire features and suppress environmental noise; (2) Dual Pool Downscale Fusion Block (DPDF) preserves multi-scale fire patterns through learnable fusion of max-average pooling outputs, mitigating small-fire detection failures. Extensive evaluation on two public datasets shows the efficient performance of our model. Our proposed model keeps the sum amount of parameters (1.45M, 51.8% lower than YOLOv8n) (4.6G, 43.2% lower than YOLOv8n), and mAP75 is higher than the mainstream real-time object detection models YOLOv8n, YOL-Ov9t, YOLOv10n, YOLO11n, YOLOv12n and other YOLOv8 variants 1.3-5.5%. For more details, please visit our repository: https://github.com/JEFfersusu/YOLO-FireAD

Figures

Figures reproduced from arXiv: 2505.20884 by the authors.

Figure 1
Figure 1. Overview of the YOLO-FireAD model architecture. 2.1 Overview As shown in [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Overview of the Attention-guided Inverted Residual Block structure. For the input features Xin ∈ R C×H×W , firstly, the channel dimension reduc￾tion is achieved by reducing the convolution to reduce the amount of calculation: Xreduce = ReLU (BN (Conv1×1(Xin))) (1) The reduction rate r = 0.25 here is experimentally optimised to make a balance between reducing information loss and computational complexity. The deep fe… view at source ↗
Figure 3
Figure 3. Overview of the convolutional additive self-attention structure. Then Xdw undergo convolutional additive self-attention[11] through the fol￾lowing process: First, the input features are decomposed into query (Q), key (K), and value (V) triples by 1×1 convolution: [Q, K, V ] = Split(Conv1×1(X)) (3) Then the parallel processing strategy of space-channel decoupling is adopted: Qˆ = Fchannel (Fspatial(Q)) (4) Kˆ = Fchan… view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Overview of the Dual Pool Downscale Fusion Block structure. For dual path feature extraction, two pooling operations are first performed in parallel to obtain complementary features: Xmax = MaxPool2×2(Xin) (9) Xavg = AvgPool2×2 (Xin) (10) The maximum pooling path reinf…
Figure 5
Figure 5. Figure 5: YOLO-FireAD iterations on the fire_detection dataset [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Comparison of model metrics (F1 score, mAP50-95, Params) on fire_detection dataset. YOLOv8n YOLO11n YOLOv12n YOLOv8n +ASF Ours [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]
Figure 7
Figure 7. Figure 7: Sample visualisation of several competitive model identification results [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]
Figure 8
Figure 8. Figure 8: Heat map visualisation results, the first row is the original image, the second row is the YOLOv8n visualisation results and the third row is the YOLO-FireAD visualisation results [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

22 extracted references · 9 canonical work pages

  1. [1]

    T. Wang, J. Wang, C. Wang, et al.,Improving YOLOX network for multi-scale fire detection, Vis. Comput., vol. 40, pp. 6493–6505, 2024. [doi: 10.1007/s00371-023- 03178-1]

    doi:10.1007/s00371-023- 2024

  2. [2]

    L. Cao, Z. Shen, and S. Xu,Efficient forest fire detection based on an improved YOLO model, Vis. Intell., vol. 2, p. 20, 2024. [doi: 10.1007/s44267-024-00053-y]

    work page doi:10.1007/s44267-024-00053-y 2024

  3. [3]

    W. Pan, X. Wang, and W. Huan,Real-time dynamic scale-aware fusion detection network: take road damage detection as an example, J. Real-Time Image Proc., vol. 22, p. 55, 2025. [doi: 10.1007/s11554-025-01634-w]

    work page doi:10.1007/s11554-025-01634-w 2025

  4. [4]

    W. Liu, Z. Shen, and S. Xu,CF-YOLO: a capable forest fire identification algo- rithm founded on YOLOv7 improvement, SIViP, vol. 18, pp. 6007–6017, 2024. [doi: 10.1007/s11760-024-03288-w]

    work page doi:10.1007/s11760-024-03288-w 2024

  5. [5]

    X. Cao, J. Wu, J. Chen, et al.,Complex Scenes Fire Object Detection Based on Fea- ture Fusion and Channel Attention, Arab J. Sci. Eng., 2024. [doi: 10.1007/s13369- 024-09471-y]

    doi:10.1007/s13369- 2024

  6. [6]

    J. Yao, J. Lei, J. Zhou, et al.,FG-YOLO: an improved YOLOv8 algorithm for real- time fire and smoke detection, SIViP, vol. 19, p. 346, 2025. [doi: 10.1007/s11760- 025-03894-2]

    doi:10.1007/s11760- 2025

  7. [7]

    W. Pan, X. Wang, and W. Huan,EFA-YOLO: An Efficient Feature Attention Model for Fire and Flame Detection, arXiv, 2024. [eprint: 2409.12635]. [Online]. Available: https://arxiv.org/abs/2409.12635

    work page Pith review arXiv 2024

  8. [8]

    Jocher, J

    G. Jocher, J. Qiu, and A. Chaurasia,Ultralytics YOLO (Version 8.0.0) [Computer software], 2023. [Online]. Available: https://github.com/ultralytics/ultralytics

    work page 2023

Show all 22 references
  1. [9]

    Zheng, P

    Z. Zheng, P. Wang, W. Liu, J. Li, R. Ye, and D. Ren, Distance-IoU Loss: Faster and Better Learning for Bounding Box Regression , Pro- ceedings of the AAAI Conference on Artificial Intelligence, vol. 34, pp. 12993–13000, 2020. [doi: 10.1609/aaai.v34i07.6999]. [Online]. Availabl...

    2020 doi

  2. [10]

    Chollet, Xception: Deep Learning with Depthwise Separable Convolutions, in 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017, pp

    F. Chollet, Xception: Deep Learning with Depthwise Separable Convolutions, in 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017, pp. 1800–1807. [doi: 10.1109/CVPR.2017.195]

    2017 doi

  3. [11]

    Zhang, L

    T. Zhang, L. Li, Y. Zhou, W. Liu, J.-N. Hwang, and X. Ji, CAS- ViT: Convolutional Additive Self-attention Vision Transformers for Efficient Mobile Applications , arXiv, 2024. [eprint: 2408.03703]. [Online]. Available: https://arxiv.org/abs/2408.03703

    2024 arXiv

  4. [12]

    Chen, S.-h

    J. Chen, S.-h. Kao, H. He, W. Zhuo, S. Wen, C.-H. Lee, and S.-H. G. Chan, Run, Don’t Walk: Chasing Higher FLOPS for Faster Neural Networks, in 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023, pp. 12021–12031. [doi: 10.1109/CVPR52729.2023.01157]

    2023

  5. [13]

    W. Chen, J. Luo, F. Zhang, et al.,A review of object detection: Datasets, per- formance evaluation, architecture, applications and current trends, Multimed. Tools Appl., vol. 83, pp. 65603–65661, 2024. [doi: 10.1007/s11042-023-17949-4]

    2024 doi

  6. [14]

    Wang, I.H

    C.Y. Wang, I.H. Yeh, and M.H.Y. Liao,YOLOv9: Learning What You Want to Learn Using Programmable Gradient Information, in Computer Vision – ECCV 2024, A. Leonardis, E. Ricci, S. Roth, O. Russakovsky, T. Sattler, G. Varol, Eds., Lecture Notes in Computer Science, vol. 15089, Spr...

    2024 doi

  7. [15]

    A. Wang, H. Chen, L. Liu, K. Chen, Z. Lin, J. Han, and G. Ding,YOLOv10: Real-Time End-to-End Object Detection , in Advances in Neural Information YOLO-FireAD 15 Processing Systems, vol. 37, 2024, pp. 107984–108011. [Online]. Available: https://proceedings.neurips.cc/paper_file...

    2024

  8. [16]

    J. Hu, L. Shen, and G. Sun,Squeeze-and-Excitation Networks, in 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2018, pp. 7132–7141. [doi: 10.1109/CVPR.2018.00745]

    2018

  9. [17]

    G. Li, Q. Fang, L. Zha, X. Gao, and N. Zheng,HAM: Hybrid attention module in deep convolutional neural networks for image classification, Pattern Recognition, vol. 129, p. 108785, 2022. [doi: 10.1016/j.patcog.2022.108785]. [Online]. Available: https://www.sciencedirect.com/sci...

    2022

  10. [18]

    Kang, C.-M

    M. Kang, C.-M. Ting, F.F. Ting, and R.C.-W. Phan, ASF-YOLO: A novel YOLO model with attentional scale sequence fusion for cell instance segmentation , Image and Vision Computing, vol. 147, p. 105057, 2024. [doi: 10.1016/j.imavis.2024.105057]. [Online]. Available: https://www.s...

    2024

  11. [19]

    Z. Tong, Y. Chen, Z. Xu, and R. Yu,Wise-IoU: Bounding Box Regression Loss with Dynamic Focusing Mechanism, arXiv, 2023. [eprint: 2301.10051]. [Online]. Available: https://arxiv.org/abs/2301.10051

    2023 arXiv

  12. [20]

    Ma and Y

    S. Ma and Y. Xu, MPDIoU: A Loss for Efficient and Accurate Bound- ing Box Regression , arXiv, 2023. [eprint: 2307.07662]. [Online]. Available: https://arxiv.org/abs/2307.07662

    2023 arXiv

  13. [21]

    C. Liu, K. Wang, Q. Li, F. Zhao, K. Zhao, and H. Ma, Powerful- IoU: More straightforward and faster bounding box regression loss with a nonmonotonic focusing mechanism , Neural Networks, vol. 170, pp. 276–284, 2024. [doi: 10.1016/j.neunet.2023.11.041]. [Online]. Available: htt...

    2024 doi

  14. [22]

    Doherty, B

    J. Doherty, B. Gardiner, E. Kerr, and N. Siddique,BiFPN-YOLO: One-stage ob- ject detection integrating Bi-Directional Feature Pyramid Networks, Pattern Recog- nition, vol. 160, p. 111209, 2025. [doi: 10.1016/j.patcog.2024.111209]. [Online]. Avail- able: https://www.sciencedire...

    2025

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.